Review on multi-stage inversion techniques of canopy reflectance models for retrieving crop variables

Abstract: Remote sensing technique is known as an inexpensive and effective tool for retrieving crop variables in a large area. The existing methodologies can be identified into two categories: the methodologies based on statistical predictive models and the methodologies based on canopy reflectance (CR) models inversion. The latter is relatively universal. Thus, it has great potential in wisdom agriculture for crop monitoring in regional scale. However, CR model inversions suffer from the so-called "ill-posed problem". Therefore, the multi-stage, sample-direction dependent, target-decisions (MSDT) inversion technique and the object-based inversion technique were previously proposed. They are similar in technical routes: the progress of an inversion is partitioned into several stages. In each stage, only a part of variables were estimated. The results of preliminary stages are used as prior knowledge of later stages of inversion. In this way, the uncertainties in parameter optimization are reduced, the ill-posed problem is therefore limited. Concretely speaking, the MSDT method firstly estimates the sensitivity and uncertainties of variables before each stage of inversion. The most sensitive and uncertain variables were firstly retrieved using a subset of remote sensing data which is sensitive to the retrieved variables. The scheme of parameterization is then updated based on the preliminary results. Another subset of sensitive variables was subsequently retrieved using another subset of sensitive data. The object-based inversion defines an "object" as a plot or a gliding window, in which the crop has similar attributes. Such attributes are referred to as "object signatures". A remotely sensed image is firstly segmented into objects. Within each object, object signatures are firstly retrieved, and used as prior knowledge in subsequent pixel-wise retrieval of spatial heterogeneous or interested variables. In this way, spatial constrains, i.e., the spatial distribution of variables, are extracted and imposed on the inversion. It can be seen the MSDT and object-based inversion essentially follow the same procedure. The major difference between them is that MSDT method makes the scheme of inversion according to the sensitivity and uncertainty of variables, while object-based inversion is based on the spatial distribution of variables. In this review, MSDT and object-based inversions were summarized into an integrated conceptual framework of "multi-stage inversion". Based on this framework, the following technical problems and the potential solutions can be summarized as follows. 1) The schemes of MSDT and object-based inversions are practically in conflict. In future studies, multi-step inversion strategies need further comparison, verification and improvement to ensure their rationality and effectiveness. The thoughts of MSDT and object-based inversions should be integrated, to develop more sophisticated inversion schemes under the conceptual framework of multi-step inversion. 2) Multi-step inversions might be significantly affected by the accuracy of preliminary parameterization of CR model. In future studies, the integrated application of multi-sources data could be helpful for CR model parameterization, and for detecting errors in each stage of inversion. For instance, same variables can be retrieved from satellite, aerial and ground remote sensing data, or obtained directly from in-situ measurements and existing remote sensing products. With approaches of scale transformation, the variables retrieved from multi-source data can be compared, in order to obtain prior-knowledge, or detect error in inversions. 3) Multi-step inversions might be distorted by error propagation. In future studies, on the one hand, gross errors and systematic errors should be detected and corrected in each stage of inversion according to the statistical distributions of retrieved variables, or by using multiple data sources. On the other hand, the schemes of multi-step parameter optimization should be customized for each variable according to its sensitivity and spatial heterogeneity. Not to fix sensitive or spatially heterogeneous variables if the accuracy and reliability of prior knowledge or the preliminary inversions could not be guaranteed.